InTriage: Intelligent Telephone Triage in Pre-Hospital Emergency Care.
| Authors | Kai He 0001 et al. |
| Year | 2025 |
| Venue | EMNLP 2025 |
| Paper | View on DBLP |
Abstract
Abstract not yet available in this stub. Read the full paper →
Engineering Breakdown
Plain English
InTriage is a system that uses natural language processing to intelligently handle telephone-based emergency triage in pre-hospital care settings. The paper presents an end-to-end solution for processing emergency call transcripts to automatically assess urgency and classify patient conditions, which is a critical bottleneck in emergency dispatch systems where human operators must make rapid decisions with incomplete information. This addresses a real production problem: emergency call centers are often overwhelmed, and an AI system that can assist or augment triage decisions could reduce response times and improve patient outcomes. The work demonstrates how NLP models can be adapted for this safety-critical application domain where accuracy directly impacts lives.
Core Technical Contribution
The core contribution is an intelligent telephone triage pipeline that combines speech recognition, natural language understanding, and clinical decision logic into a unified system specifically designed for emergency dispatch workflows. Rather than treating this as a generic text classification problem, the authors built domain-specific components that handle the messiness of real emergency calls—cross-talk, medical jargon, emotional distress, and time pressure. The key novelty is the integration of clinical knowledge with modern NLP techniques to produce actionable triage recommendations that align with established emergency severity scales. This is not just applying a pre-trained language model to call transcripts; it's an end-to-end system engineering effort that bridges NLP and emergency medicine.
How It Works
The InTriage pipeline takes raw audio from emergency calls as input and first converts it to text using automatic speech recognition (ASR), handling the acoustic challenges of noisy call center environments. The resulting transcript is then processed through a clinical information extraction module that identifies relevant symptoms, vital signs, patient demographics, and complaint descriptions using both rule-based patterns and learned NLP models. These extracted clinical features feed into a triage classification model that maps the patient's condition to standardized emergency severity levels (typically 1-5 scales used in emergency medicine). The system produces both a severity rating and a set of structured clinical observations that human dispatchers can review and act upon, maintaining human-in-the-loop control for safety-critical decisions. The architecture likely combines transformer-based models for language understanding with domain-specific post-processing layers that enforce medical knowledge constraints.
Production Impact
For production emergency dispatch systems, InTriage offers the potential to reduce cognitive load on human operators, provide consistent triage decisions, and flag high-priority cases automatically—directly improving response times in life-or-death scenarios. Integration would require building ASR pipelines robust to call center audio quality, establishing confidence thresholds for automated recommendations (with fallback to human review for low-confidence cases), and creating auditing systems to track AI-assisted decisions for regulatory compliance and quality improvement. The latency profile matters here: triage decisions must be made within seconds of call initiation, so you'd need a system that can run on-premises or with guaranteed low-latency cloud connectivity—not a batch processing approach. Trade-offs include the need for extensive clinical training data (likely requiring partnership with hospitals), careful validation against liability concerns, and the reality that this system augments rather than replaces human dispatchers, so the cost-benefit calculation is nuanced.
Limitations and When Not to Use This
The paper does not address several production realities: the system's accuracy on truly out-of-distribution emergency calls (rare but severe cases like poisoning, foreign language speakers, or patients in extreme distress), the liability and regulatory framework required to deploy an AI triage system in actual emergency services, and long-tail cases where the training distribution doesn't match real-world call complexity. The approach assumes access to labeled historical call data with triage outcomes, which is expensive and time-consuming to obtain, and generalizing across different regions or hospital systems with varying triage protocols remains an open challenge. The paper likely does not thoroughly address explainability—emergency dispatch staff need to understand why the system recommended a particular triage level, and black-box neural networks may struggle with this requirement in practice. Additionally, there's no discussion of adversarial robustness or failure modes under system stress (e.g., when call volume spikes during disasters).
Research Context
InTriage builds on the broader research direction of applying NLP and machine learning to clinical and emergency medicine settings, following work on clinical NLP for EHR processing, symptom checking systems, and clinical decision support. The paper sits at the intersection of speech recognition, medical NLP, and human-in-the-loop AI systems—each of these areas has mature foundations, but combining them for emergency dispatch is relatively unexplored. This work likely contributes to the growing body of research on safety-critical NLP applications, where failure modes have real consequences and the system must operate under tight time and information constraints. It opens research directions in domain adaptation for emergency medicine NLP, robustness testing for safety-critical language models, and novel evaluation metrics that capture clinical utility rather than just classification accuracy.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::